A machine-learning virtual source model for nanoscale transistors.

Area of Science:

• Semiconductor device physics
• Machine learning applications in electronics

Background:

• Purely data-driven neural networks (NNs) for transistor modeling face challenges with data requirements, physical accuracy, and generalization.
• Conventional virtual source (VS) models require manual parameter extraction, limiting their efficiency.

Purpose of the Study:

• To develop an integrated neural network-virtual source (NN-VS) model combining physical principles with machine learning.
• To enhance accuracy, data efficiency, and extrapolation capabilities in transistor modeling.

Main Methods:

• Developed an end-to-end trainable hybrid model integrating NN adaptive learning with the VS framework.
• Validated the model on experimental nanoscale transistor data, including 2D semiconductor FETs and silicon FinFETs.

Main Results:

• The NN-VS model significantly outperforms pure NN models in accuracy, data efficiency, and extrapolation.
• Demonstrated robust performance even in data-limited scenarios.
• Successfully captured experimental characteristics of various nanoscale transistors.

Conclusions:

• The NN-VS model offers a superior alternative to traditional and purely data-driven approaches for transistor modeling.
• This hybrid model enhances the reliability and applicability of NNs in semiconductor device analysis and circuit simulation, particularly for emerging technologies like ferroelectric FETs.